Multi-scale dual-discriminator GAN completes occluded plant point clouds

Researchers have developed a multi-scale dual-discriminator generative adversarial network (MSDDG) to complete missing point clouds of plants, a common problem when optical sensors capture complex canopies. The method combines a multi-resolution encoder (MRE), a multi-scale point cloud generator (MSPG), and two discriminators—one judging three-dimensional (3D) spatial distance and one judging multi-view projected silhouettes. Trained on Plant4L, a new dataset covering sunflower, pumpkin, luffa and eggplant, the framework reconstructs occluded leaves and stems with high geometric fidelity. It also supports single-view 3D reconstruction, improving leaf-area and inclination-angle estimates. The work addresses a key bottleneck in plant phenotyping: incomplete 3D data caused by viewpoint occlusion and sensor limitations.

Three-dimensional plant reconstruction is central to crop phenotyping, growth monitoring and precision agriculture. Optical imaging, multi-view stereo, structure from motion, and point cloud registration can capture plant architecture, but viewpoint occlusion and sensor limits often leave large gaps, especially inside dense canopies. Voxel-based completion is computationally heavy; deep-learning methods may lose local detail or global consistency; projection-constrained methods depend on precise image-point cloud alignment and view selection. These gaps reduce the reliability of downstream traits such as leaf area, angle, and organ topology. Based on these challenges, in-depth research is needed to develop completion methods that preserve both fine plant geometry and overall structural realism.

A team from the School of Automation at Hangzhou Dianzi University and the Jiangsu Academy of Agricultural Sciences Wuxi Branch published ((DOI: 10.1016/j.plaphe.2026.100218) the study in Plant Phenomics in 2026. The article, Digital Object Identifier, introduces MSDDG, a dual-discriminator adversarial framework for plant point cloud completion. It also presents Plant4L, a dataset built from four crop species and augmented with non-rigid transformations and virtual occlusion. The network predicts missing regions from visible point clouds and uses multi-view silhouettes as training constraints. The work was supported by the National Key Research and Development Program of China and Zhejiang provincial funding programs.

Multi-scale dual-discriminator generative adversarial network (MSDDG) is built around three linked ideas. First, a multi-resolution encoder (MRE) based on cascaded multilayer perceptrons extracts features at several scales and concatenates them into a 1,920-dimensional latent vector, preserving global architecture and local organ detail. Second, the multi-scale point cloud generator (MSPG) decodes this vector through hierarchical fully connected layers and skip connections, producing coarse-to-fine predictions of the missing region. Third, two discriminators provide complementary supervision: a spatial distance discriminator evaluates whether generated points follow realistic three-dimensional (3D) distributions, while a multi-view projected silhouette discriminator compares two-dimensional (2D) contours rendered from the generated cloud with ground-truth silhouettes. A virtual camera system creates visible and occluded regions during training, and density-aware radius estimation fills projection holes. On Plant4L, which contains 1,900 samples across sunflower, pumpkin, eggplant and luffa, MSDDG achieved average Chamfer Distance (CD), Hausdorff Distance (HD) and Uniformity Chamfer Distance (UCD) below 0.41, and an F1 score of 0.906. In single-view tests, completion raised leaf-area R² from 0.8441 to 0.9654 and leaf-inclination R² from 0.9092 to 0.9743. Spatially concentrated gaps are recovered best; highly scattered occlusions remain harder because the generator outputs a fixed 512 points.

The authors said the dual-discriminator design was central to the advance. They explained that one discriminator keeps the generated cloud plausible in 3D space, while the other checks whether its projected outlines match real plant silhouettes. This combination, they said, prevents the model from sacrificing structural correctness just to satisfy a single view. They also said the Plant4L dataset gives the network diverse but biologically grounded examples, helping it learn shared plant geometry rather than memorizing one crop. According to the authors, the framework is especially useful when occlusion creates large, continuous gaps in field or single-view imaging.

MSDDG could support plant phenotyping, digital modeling and ecological monitoring by turning incomplete single-view scans into more usable 3D structures. In tests, the completed clouds improved leaf-area and inclination-angle estimates, two traits that influence light interception, canopy architecture and yield-related analysis. The approach may reduce reliance on multiple scans and complex registration, which is valuable for field phenotyping where time, space and sensor access are limited. The authors caution that performance declines for unseen plant types, extremely sparse data or widely scattered missing regions. Future work may add adaptive point generation, lightweight segmentation and structural priors linking stems and leaves to improve generalization and efficiency.

###

References

DOI

10.1016/j.plaphe.2026.100218

Original Source URL

https://doi.org/10.1016/j.plaphe.2026.100218

Funding information

This research was supported by National Key R&D Program of China (2022YFD2000100), Pioneer and Leading Goose R&D Program of Zhejiang (2022C02026), Zhejiang Provincial Natural Science Foundation of China under Grant No. (LTGY23F030001).

About Plant Phenomics

Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting‑edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real‑world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high‑throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.